I'm an AI Software Engineer building agentic systems, LLM infrastructure, and applied AI products. I worked with multi-agent orchestration, real-time voice pipelines, fine-tuning, and production observability. I also bring solid backend depth in .NET 8, C#, and enterprise software engineering from prior roles building production systems. I care about clean abstractions, real-world reliability, and shipping things that work end-to-end.

Open to AI Engineer, AI Software Engineer, Applied AI, and Forward Deployed roles, especially places shipping agentic systems with real users.

• Agentic Systems — multi-agent orchestration, LLM reasoning loops, autonomous recovery
• Real-Time Voice AI — streaming STT/TTS, VAD, barge-in handling, low-latency loops
• Applied AI & LLM Infra — RAG pipelines, fine-tuning (QLoRA/PEFT), evaluation frameworks
• Backend & APIs — FastAPI, .NET 8, async workflows, modular architectures
• Observability & MLOps — OpenTelemetry, Langfuse, W&B, experiment tracking

Solo-built, live full-stack dashboard for comparing LLM cost, latency, and infrastructure. Three modules: Model Arena (LMSYS ELO + OpenRouter), Cost Calculator, and Infra Explorer across 15 providers and 10 GPUs. Includes a Groq Llama 3.3 70B advisor, rate limiting, and prompt injection defenses. React 18 Vite Zustand SWR Tailwind Recharts Groq Vercel

Autonomous self-healing agent for decentralized cloud. Uses Llama 3.3 70B to diagnose deployment failures, selects recovery actions via confidence scoring, and redeploys workloads across Akash providers. Decision traces via Langfuse. Python FastAPI Akash Network Llama 3.3 70B Langfuse Telegram

Multi-label emotion classification over 1M+ tweets. Benchmarked RoBERTa, DistilBERT, and ALBERT before fine-tuning Qwen3-0.6B with 4-bit QLoRA. Macro-F1 0.74 → 0.81 via per-class threshold optimization and synonym augmentation. Tracked with W&B, deployed on Hugging Face Spaces. PyTorch Qwen3 QLoRA HuggingFace W&B

AI-powered due diligence copilot for VC investors. Upload a pitch deck, get confidence-scored analysis with verified claims and a voice-narrated deal memo. Claim-level routing across graph, web, and LLM verifiers with OpenTelemetry tracing. Python FastAPI Strands Agents Next.js Memgraph OpenTelemetry D3.js

Real-time multimodal voice agent work: streaming transcription, voice activity detection, and low-latency LLM reasoning loops with barge-in handling. Focused on the reliability problems specific to live audio, including turn-taking, interruption, and hallucination control. Repos private. FastAPI Streaming STT VAD LLM Orchestration Langfuse

More on GitHub — GlassesToReels, NutriGuard (RAG with Ragas evals), and others.

Languages

AI / ML / Agents

Backend & Frontend

Observability & MLOps

Cloud, Data & DevOps

I'm open to AI Engineer, AI Software Engineer, Applied AI, and Forward Deployed roles, especially places shipping agentic systems with real users.

📬 LinkedIn · 🌐 Portfolio